A Leap Forward in Storage Efficiency with the OFP Initiative and Hammerspace

My name's Kurt Kine. I'm with Hammer Space and I am the senior director of AI marketing. And today I'll be giving you an update on the Open Flash platform initiative.

Um, so here we are, hammer Space, a software company, but we're at AI Infrastructure Day, so it wouldn't be AI Infrastructure Day without a little hands-on show and tell. So I brought with me our prototype iteration of the open flash platform reference design, and I'll pass this around for folks in the room while I'm, uh, chatting about it. Um, you'd be aware it is a prototype, so there's a few things sticking out but fairly robust.

Um, so a quick update on, Hey Kurt, how much flash is this? What we're doing? So this is our initial U2 concept.

Um, and so each flash drive I think is currently, I can read off the top what, what size those are. 62 B. Yeah, so these are, these are smaller drives, but you know, we can go to the largest, um, U2 form factor right now and there's eight of them in there.

Um, and so what is the open flash platform, right? It's an initiative that the collection of vendors that are striving to really drive down the complexity within storage as well as, you know, the cost. Um, and cost is, you know, a multi-variable factor.

Um, and we think open standards and commodity-based hardware will really get us there, but it is gonna take some design iterations. Uh, and our goals, right, are essentially to be able to achieve one exabyte in a rack from this open architecture. Now we know deployment wise customers are not necessarily going to deploy an exabyte rack.

Um, so we actually see these more as a way to disappear your storage into your data center, as in most racks have one U of face in them where they're using blanking panels. And if you've got something that's consuming a nominal amount of power and producing a nominal amount of heat, you've probably got room for it. And so you can distribute these throughout a data center and really hide your storage.

And when we get to this really incredible capacity, right, it's a really a huge amount of capacity that you can distribute through the data center. Now, if you just step things that we've gone through as we've iterated, um, you know, early on we looked at a to you form factor, um, and we figured out, you know, there's a ton of challenges that come with that. Um, you know, things like the, um, chassis itself starts deforming with weight.

Um, you start running into some other things around, um, power and cooling and you create wasted space with two U. Um, so we also wanted to increase the granularity of the system, right? And so we hit on this one U form factor, um, and essentially right five of those fit into a standard, um, rack space did, took a ton of right temperature analysis to be able to make sure we can run all of these components in this really, really tight space.

And that's what we're striving to get through, is continue to increase the density of the system, right? We're, um, running pressure analysis through these things. We're looking at the flow speed.

I mean, it's taking iteration after iteration for us to really be able to achieve the density that we're looking for. And as I mentioned when I first started this, passing this around, um, this is our initial right reference design. Um, and we're currently working out all of the prototyping work that goes with that.

Um, and one of the big breakthroughs that we made, um, was in selecting the right chip to be able to run this thing right? Um, and we had looked at a number of dpu, uh, and we have a number of DPU vendors that have, um, you know, suggested designs. Um, but when we came across the X site chip, um, we got really excited.

Now, I think it's pretty amazing they, um, take an approach, uh, that is, you know, actually from the compute side, let's put as much compute into the system as possible within an envelope as opposed to, Hey, we have this nick and let's add a little bit of feature set on it. So what we get is essentially a computing chip that maybe three to five years ago was what you got from an X 86 server, but in this really amazing, um, power efficient package and it lets us run, you know, Linux on it and be able to serve storage as powerful as a storage server within this really tight design. So Ted, I wanted to, um, invite you up just really quick to talk about, um, the X site, uh, E one chip.

Thanks Kurt. Yeah, yeah. Great.

So, um, we partnered, uh, through the open flash platform together, uh, in order to bring the densest warm flash platform to market. Um, and we're excited about it because our particular DPU is the world's first 800 gig DPU. We sampled it may of last year.

And our nearest competitor, which is Nvidia, we'll have a Bluefield four, uh, we think sampling Q3 this year. We have a solid year plus start on that. What makes our product exciting is exactly what you said.

We've got 64 neo versus two arm cores there, and then a programmable nick also inside. That's what makes it a DPU with two by 400 gig or eight by 100 gig 30 on it. It's a five nanometer product that in this application will only burn 75 watts.

So it's perfectly said, it's, it's like getting a server from a couple years ago that'll fit in a space like this. Um, so that's really it. It's, it's we're integrated at a very aggressive level.

Um, and there's a programming model that's different as well. Uh, DPU tend to have a proprietary nick pipeline that you have to code and it's its own little operating system off to the side, and then there's a choke point into the arm course. They're typically arm cores.

Um, so we have a lot of arm cores, you know, two x the competition that's out now, uh, and the same number as the leading competitor will be out here in a couple quarters. Um, but we don't have any choke point between the performance nick side, if you want to call it that, or the, we call it the ethernet unit and the actual arm complex. So it's a all fast path design and this turns heads because people assume that if you're riding on top of a standard Linux model, use things like DPDK and your tool flow that you're just not gonna have line rate.

So we came out a year early, maybe a year and a half early, uh, and we just proved, uh, that we could run it 800 gig line rate. Um, and we did that independently with Keysight with a test called Dash Hero 800 that Microsoft sort of founded with their sonic group. Um, so enough said for now, thank you for being our, our deep partner and we are building that system together and more on that, uh, later, uh, 'cause I've got a tech field day coming up.

Yeah, Stay, stay online and uh, join us for, uh, Ted's Yeah. Tech field day. I know we are over time now.

Um, I'm just gonna run through, um, just one more thing here. Um, so, you know, we've talked about the fact that there's a flash shortage right now. Um, and we were looking at a potential future standard, um, to be able to increase the density even more to get to that one exabyte rack.

Uh, we still think that is in plan. Uh, we're still talking to the flash vendors about this new form factor. Um, we've found limitations with the current E two farm factor that we'll need to be able to go beyond.

Um, and those discussions are really progressing. Um, so you see some just great industry participation, um, in this OFP initiative. We're starting to, um, do the paperwork and everything to become part of OCP, right?

We definitely wanna push this into as a standard form factor. Uh, and we want, the reason why it's truly open, again, goes off a lot back to the um, chip that we chose, right? It's not just for necessarily, um, the shared storage vendors to use, right?

This, it really is a fully our server that's incredibly efficient. So yes, scalable file storage for all these large file systems, but we're talking to other customers about the use case for block storage and these hyperscalers who have all sorts of different flash devices in their data center, in their servers all over the place, right? Would love a homogeneous boot device that would be able to support every server within a rack.

So essentially, you know, we've got top of rack switching, well now we have top of rack storage and you would be able to boot off this single type of system, use that as your boot drives for all the servers in rack and eliminate a lot of qualification time. Um, eliminate a lot of this problems with heterogeneous, um, you know, devices spread throughout the data center. You can choose one type, qualify that, and now every server and potentially boot off of that single design.

Here's the timeline. We are currently in prototyping and entering validation. We've got two early access customers that we're looking to ship to this quarter.

We're super excited to get their feedback and then iterate on this design. Um, and we're still shooting for general availability of this in the second half of the year. And as always, we are also continuing to recruit other folks into this initiative, including other storage vendors.

We really wanna drive this as an industry standard, um, so that, you know, it's adoptable by everyone, um, and that everyone can benefit from this efficiency, um, and flash capacity. So thanks for your time.